Method and system for FFT-based companding for automatic speech recognition

1. A method for processing a speech signal, comprising the steps of: performing a fast Fourier transform on a speech signal to produce a speech signal having a plurality of frequency bands in a frequency domain, and for each frequency bands further comprising the steps of: filtering the speech signal in the frequency domain with a spatial broadband filter; compressing the broadband filtered speech signal; filtering the compressed speech signal with a spatial narrowband filter; and expanding the narrowband filtered signal to an expanded signal.

2. The method of claim 1 , in which the speech signal includes noise of an environment in which the speech signal is acquired.

3. The method of claim 1 , in which the speech signal is sampled into a plurality of overlapping frames, and the broadband filtering, compressing, narrowband filtering, and expanding is performed individually on the plurality of frames.

4. The method of claim 1 , further comprising: summing the expanded signals into a summed speech signal; and recognizing the summed speech signal by an automatic speech recognizer.

5. The method of claim 1 , in which the plurality of frequency bands are equally spaced.

6. The method of claim 1 , in which cepstral features are determined directly from the expanded signal.

7. The method of claim 1 , in which the broadband filter and the narrowband filter are linear.

8. The method of claim 1 , in which a Fourier spectrum of a response of the broadband filter is a vector F, and a spectrum of the broadband filtered signal is X 1 =F{circle around (×)}X, where {circle around (×)} represents an element-wise Hadamard multiplication, and X is the speech signal, and X 1 is the broad band filtered signal, and in which the expanding produces the expanded signal X 2 =|X 1 | n-1 X 1 , where n is a constant companding factor.